A language model has to turn text into something it can process, and there are three broad answers: subword tokens, raw bytes, or images of rendered text. Comparisons between them are usually unfair in a way that is hard to see, because each encoding exposes a different amount of actual linguistic content depending on the language.

That unfairness is not evenly distributed. Tokenisers trained mostly on English chop other scripts into far more pieces, so a model reading Hindi or Amharic sees a smaller slice of the sentence within the same budget. Comparing encodings without controlling for this measures the tokeniser's training data as much as the encoding itself.

The study controls both content and downstream capacity, using verified parallel sentences across thirteen languages and five scripts, so each encoding is judged on what it preserves rather than on how much it happened to be handed. Framing it as a rate-utility frontier is apt: the question was never which encoding is best, but what each costs for what it keeps.